AI governance: how to keep an AI agent registry

Controlling artificial intelligence starts with a rigorous inventory. An agent registry acts as the operational backbone of any system, detailing each asset precisely: from the human owner and the data involved to the limits of authority, the performance KPIs and the shutdown protocol. Without that file you are operating blind, with no way to evidence regulatory fit, no technical audit and no capacity to respond to a critical failure. Note: this content is purely informational. Check your legal obligations against Regulation (EU) 2024/1689, the European Commission and your national supervisory and data protection authorities.

In short

• AI governance is not regulatory compliance: it is knowing which agents exist, what they can do and who answers for them.
• The agent registry is governance's first deliverable and the basis of any later audit.
• Each agent needs a named human owner, not a department.
• An agent with no withdrawal plan is technical debt with permissions.
• The Governance axis is worth 20 of the 100 points of the IMAN (AI-Native Maturity Index).

What is AI governance and how does it differ from compliance?

Compliance answers "is this legal?". Governance answers "is this under control?". They are different, and the second is a condition of the first: you cannot evidence that a system meets the rules if you do not know it exists.

In practice, AI governance covers four things:

1. Inventory: which AI agents and systems exist, including those a department contracted without going through IT.
2. Accountability: who answers for each one to the board.
3. Limits: what each agent can do and how far, in numbers.
4. Evidence: what gets logged so it can be demonstrated afterwards.

Why is the agent registry the first deliverable?

Because it resolves three problems companies usually tackle separately: security risk, regulatory requirement and runaway cost.

A common example: a services company of 120 employees discovers in its first mapping exercise nine active AI tools, seven contracted outside IT, three with access to corporate email and none with an assigned owner. From our experience on NEXTAI projects, that finding is the rule and not the exception in a first diagnosis.

The registry turns that scatter into a table with owner, purpose and permissions. From there you can decide; before, you cannot.

It should not be confused with the GDPR record of processing activities. They are different, complementary documents: the processing record looks at personal data and its purposes; the agent registry looks at capabilities to act and their limits. The same agent appears in both, with different information. We cover that in GDPR and artificial intelligence.

Nor is it a paperwork exercise. The registry is what lets you answer, in minutes, three questions that always arrive at the worst moment: which agent did this, who answers for it and how is it switched off.

What fields should an agent record have? (full template)

This is the template NEXTAI delivers in phase 5 of the NEXT-5 Roadmap. One per agent, and it is versioned.

# AGENT RECORD — AI agent registry
identifier: AG-014
name: "Accounting Superagent"
version: "2.1"
status: production            # design | pilot | production | suspended | retired
created: 2026-03-10
last_review: 2026-08-15
next_review: 2026-11-15

owner:
  business_owner: "Full name — Head of Accounting"
  technical_owner: "Full name — IT"
  deputy: "Full name"

purpose:
  process_owned: "Posting purchase invoices for the client portfolio"
  problem_solved: "6,200 invoices/month posted by hand"
  out_of_scope: "Year-end tax close, payroll, client advice"

data:
  read_sources: ["invoices@ mailbox", "Document manager", "ERP: master data and entries"]
  write_destinations: ["ERP: journal entries in draft status"]
  personal_data: yes
  special_categories: no
  legal_basis: "Performance of the contract with the client"
  processing_region: "EU"
  retention: "Prompts and outputs 90 days; action log 5 years"

permissions:
  technical_user: "svc-ag014"
  scope: "Read only except creating a draft journal entry"
  authorised_tools: ["ocr_document", "query_master_data", "create_draft_entry"]
  allowed_network_domains: ["erp.internal", "api.ocr-provider"]
  explicit_prohibitions: ["Modify master data", "Send external email", "Post entries definitively"]

limits_of_authority:
  max_amount_per_document: "€25,000"
  max_volume_per_hour: 400
  confidence_threshold: 0.85          # below this, it goes to human review
  actions_requiring_human_approval: ["New supplier", "Intra-EU transaction", "Reverse charge"]
  operating_hours: "Mon-Fri 06:00-22:00"
  monthly_budget: "€600"

kpis:
  target_rao: "65%"
  current_rao: "63%"
  material_error_rate: "< 0.5%"
  cost_per_document: "€0.07"
  average_time_per_document: "11 s"
  sampling_review: "5% weekly"

risk_and_regulation:
  use_classification: "Minimal risk — pending legal validation"
  dpia_required: no
  open_incidents: 0

withdrawal_plan:
  triggers: ["Material error > 2% in a month", "ERP change", "Cost > €900/month", "Business decision"]
  procedure: ["Move to the human approval band", "Suspend the technical user", "Revoke credentials and tokens", "Export the action log", "Reassign the process to the human team", "Archive the record with the reason"]
  execution_owner: "Technical owner"
  target_time: "4 hours"
  dependencies_to_review: ["AG-011 Document Intake"]

The eight fields that can never be left blank are: owner, purpose, data accessed, permissions, limits of authority, KPIs, withdrawal plan and review date. An agent with any of those blank should not go into production.

Who governs what?

The committee does not need to be heavy. In a company of 100–250 employees, a quarterly one-hour meeting with the registry in front of you is usually enough. What does not work is diluting accountability: an agent whose owner is "the admin department" is, in practice, an agent with no owner.

Which governance KPIs should be reviewed each quarter?

1. Number of active agents and the change on the previous quarter.
2. Agents with no assigned owner: the target is zero.
3. Records with an overdue review: the target is zero.
4. RAO (Operational Autonomy Ratio) per process and the gap against target.
5. Material error rate per agent.
6. Monthly cost per agent against budget.
7. Security incidents and time to resolution.
8. Agents withdrawn and the reason, which is the indicator the organisation learns most from.

How do you roll out the registry in 90 days?

1. Weeks 1–2. Open inventory: ask each department which AI tools it uses, with no penalty for what turns up.
2. Weeks 2–4. A minimum record for each item in the inventory, even if it is not an in-house agent.
3. Weeks 4–6. Assign an owner and a technical owner to each record.
4. Weeks 6–9. Limits of authority and KPIs for the agents in production.
5. Weeks 9–12. Withdrawal plan, first committee review and the quarterly calendar.

How NEXTAI does it

NEXTAI delivers the agent registry as part of phase 5 of the NEXT-5 Roadmap, built into the Governance layer of the Enterprise Brain. No Superagent is deployed without a complete record, a human owner, numeric limits and a withdrawal plan. The result shows up in the Governance axis of the IMAN (AI-Native Maturity Index), which carries 20 of 100 points.

Frequently asked questions

Is an AI agent registry mandatory? The specific obligation depends on the framework applicable to each company and on the use of each system; check with your legal adviser. Regardless of that, it is the practical prerequisite for any audit: you cannot demonstrate control of what is not inventoried.

Does a spreadsheet work as a registry? At the start, yes, and it beats having nothing. It stops working once there are more than ten agents, or once the registry has to be consulted from several departments, because a spreadsheet has no version control, no permissions and no trace of who changed what.

What is the difference between an agent owner and a technical owner? The owner answers for the agent doing its job well and meeting its KPIs; that is a business role. The technical owner answers for permissions, integrations and being able to switch it off. Separating the two roles stops a business problem being treated as an IT ticket.

When should an agent be withdrawn? When one of the thresholds defined in its record is triggered: material error above the limit, cost overrun, a change in the system it integrates with, or simply that the process no longer exists. A planned withdrawal is a sign of maturity, not of failure.

Each agent record needs a quarterly update in production environments, plus immediate adjustment whenever the process, the system architecture or the applicable legal framework changes. Making the next review date a mandatory field removes the dependence on human memory. If you want to systematise the agent inventory in your organisation, request our free Digital Audit to start the process.